AI2U Review
A 3D dating sim that bolts a real-time facial-animation engine onto an LLM chat backend. The first AI girlfriend product that actually feels like a game - for better and worse.
Verdict: a 3D dating sim, not a chat app - judge it that way
AI2U is the only product on our shortlist that ships through Steam, installs to your hard drive, and renders a character in real-time 3D. Inside its own category - LLM-driven visual novel - it's the closest thing to a finished commercial product anyone's made. Outside that category, the comparisons get unfair: the dialogue engine is GPT-3.5-class, the save files have a history of breaking on engine patches, and you need an actual GPU. Buy it if you came for a game; do not buy it expecting Candy with a 3D model.
What AI2U actually is
AI2U is a 3D dating sim that shipped on Steam in late 2024 from a small Japanese-American studio. It is not an app. It downloads to your hard drive at ~24 GB, runs on Unity, and ties a real-time facial-animation system to a cloud LLM that writes the dialogue live. Closer in shape to VA-11 Hall-A or a Persona social-link scene than to anything in our usual chat-app shortlist.
The trick the studio is selling: you don't pick dialogue from a wheel. You type. Or you talk into a mic. The model replies in character, the avatar reacts with expression blends keyed to the sentiment of her line, and a story scaffold (eight main characters, three ending paths each) gives the whole thing a shape that ends. Between sessions, your chosen character "lives" in a small apartment scene you can drop back into.
The loop, in detail
Pick one of eight main characters at the start of a save and the game opens on a meet-cute. From there the structure stays roughly the same across routes:
- Scripted scenes. First date at the café, beach day, rainy-night apartment. Backgrounds and blocking are pre-built; dialogue is generated live and voice-acted. Each main character has roughly 14–18 of these.
- Free-chat hubs. Sit on the couch, in the kitchen, at the corner bar - and just talk. Idle animations, head turns, contextual expressions all run in real time. This is where the LLM does its heaviest lifting.
- A hidden affection arc. No visible meter, but the model tracks what you've said and the writing team encodes "milestones" that gate new scenes. You feel the gate trip when a new location unlocks.
- Between-session events. Text messages arrive a few hours after you quit. Voice notes drop overnight. Skip a few days and you might come back to a sulking message or a surprise outing scene queued up.
The pacing is deliberately slow. A good-ending run averages 15–25 hours of play, spread across two or three real-world weeks - some scenes only fire after enough wall-clock time has passed since the last one. Visual-novel veterans will read that as intentional. Players who wanted a weekend binge will read it as a chore. Both are right.
Cast & storylines
The base game ships with eight characters - four feminine, four masculine. Each has a credited voice actor, a backstory bible, and roughly 20 hours of unique content. The writing isn't Persona 5, but it sits cleanly above most visual novels, and the studio invested in the kind of small consistency details (running jokes, specific drink orders, mannerisms) that survive LLM generation.
Eight base routes
Standouts: Maya, the cynical bartender; Lin, the anxious grad student; Theo, the jaded chef trying to reopen his late mother's restaurant; Riku, the cocky musician with stage fright. Each route is ~20 hours.
Three endings per character
Which one you unlock isn't a dialogue-wheel pick - it's a function of the affection arc and how you actually spoke to her in free-chat hubs. Reroll-proof in a satisfying way.
Full voice acting in scripted scenes
Every set-piece is fully voice-acted, with a named cast in the credits. Free-chat lines use a real-time TTS tuned to each actor's tone - close enough to hold the illusion, but you can hear the seam.
Real-time facial blends
The studio's expression system reads sentiment from the LLM response and drives a custom face rig in real time. The first time a character genuinely winces at something you typed, it lands harder than any chat-app screenshot ever has.
Five DLC character packs so far
$5–$10 each. Quality is uneven - the "Aria" pack (psych student, ten scenes) is the best post-launch addition; the "Holiday" pack feels phoned-in. Buy individually, not as a season pass.
Photo mode
Pause anywhere, fly the camera, pose the character, change lighting, swap lenses. The modding community has used it to cut short films. The depth here surprised us.
The dialogue model - what it can and can't do
This is the honest weak spot. AI2U's dialogue engine is GPT-3.5-class, not GPT-4 or Claude Sonnet. The studio confirmed this in a 2025 dev blog: because the base purchase includes a generous monthly token budget, per-user cost has to stay low, and 3.5-class is the model that survives the math. The practical effect is responses that are coherent and in-character - but noticeably shallower than what you'd get from Candy AI for the same input.
Our 200-message coherence run scored AI2U at 38/50. Solid for a game; a step behind dedicated chat apps. The texture:
- Character lock is the engine's strongest move. Writers gave each character a tight voice bible, and the personality preset holds it. Maya keeps her sarcasm 40 messages in. Lin keeps her anxious hedging. The lock is firmer than raw ChatGPT with the same prompt.
- Memory is medium. Within a save, she remembers major beats - your job, the argument from chapter 2, the fact you missed her birthday scene. Across saves, full wipe. A 2025 patch added opt-in cloud-saved memory packets, but it's off by default and the studio has been quiet about it since.
- Emotional nuance is the GPT-3.5 tell. Subtle turns sometimes get flattened. She may pivot from a serious confession to "anyway, what should we have for dinner" without the bridge a GPT-4 model would write.
- Failure mode is out-of-character drift when you push toward subjects outside her domain - politics, recent news, technical questions. The model defaults to a vaguely ChatGPT-ish register and the illusion slips. Stay in-world and it holds.
%APPDATA%\AI2U\Saves on Windows or ~/Library/Application Support/AI2U on macOS before any patch. This is the single most-discussed complaint in Steam reviews.
System requirements (and what they actually mean)
AI2U renders a rigged 3D character in real time and ships dialogue calls over the network. That's a real GPU job - not "any laptop will do." Integrated Intel UHD and AMD Vega iGPUs fail the character-load step. The Steam page lists conservative minimums; we ran our own tests and the table below reflects what actually works.
| Spec | Minimum | Recommended | Our notes |
|---|---|---|---|
| OS | Windows 10 64-bit | Windows 11 | macOS works via Apple Silicon native build; Linux via Proton (community-supported). |
| CPU | Intel i5-8400 / Ryzen 5 2600 | i5-12400 / Ryzen 5 5600 | CPU is mostly idle - GPU is the bottleneck. |
| RAM | 8 GB | 16 GB | 8 GB works, but 16 GB removes hitches. |
| GPU | GTX 1060 / RX 580 (4 GB VRAM) | RTX 3060 / RX 6600 (8 GB VRAM) | VRAM is the hard requirement. Below 4 GB, characters won't load. |
| Storage | 24 GB | 40 GB (with DLC) | SSD strongly recommended - scene loads are noticeably faster. |
| Network | Broadband | Broadband, low latency | Dialogue calls go to the cloud - high ping causes awkward pauses. |
The Mac build is a native Apple Silicon binary, not a Rosetta translation. On an M2 Air (16 GB), it held a stable 60 fps at medium settings through 8 hours of testing. No Linux native build - Proton works for most users via community-supported reports. No console release planned and no GeForce Now / Boosteroid availability as of June 2026.
Pricing & DLC - what to actually buy
Base game
- 8 main characters
- All story routes
- Voice acting included
- Generous monthly token budget
- Free patches indefinitely
Deluxe Edition
- Base game
- 3 DLC character packs
- Soundtrack
- 2× monthly token budget
- Photo-mode prop pack
- Early access to new DLC
Token top-up
- Optional - only if you chat heavily
- Unlimited dialogue tokens
- No effect on story
- Cancel anytime
What we'd recommend: base game at $19.99 to find out if you like it; if you do, add the Aria DLC ($8) before the Deluxe bundle, which front-loads two weaker packs. Token top-up is only worth $4.99/mo if you're chatting more than ~3 hours a day in free-chat hubs - most players never hit the included budget. The adult patch is free, but it's distributed off-Steam and requires an age-verified account on the studio's site (Steam doesn't permit on-platform adult content for this product category in most regions). Prices verified May 2026.
Pros & cons
What works
- Real-time expression blends keyed to LLM sentiment - the first AI companion product where the character physically reacts
- Every scripted scene fully voice-acted by a credited cast
- $19.99 one-time purchase model - pay once, own forever, no $15/month forever
- Eight base routes with distinct, well-written voice bibles that survive LLM generation
- Photo mode deep enough that modders cut short films with it
- Native Apple Silicon build runs stable 60fps medium on an M2 Air
Where the 3.9 ceiling lives
- Dialogue engine is GPT-3.5-class - noticeably shallower than Candy or Anima on the same prompt
- Hard requirement of 4 GB+ VRAM dedicated GPU - laptops with iGPUs simply fail the character-load step
- Saves have broken on two engine updates; the migration tool helps but doesn't fully fix it
- PC only - no mobile, no console, no cloud streaming as of June 2026
- Small modding scene, slow patch cadence (~3 content updates per year)
- Adult patch lives off-Steam and requires age-verifying on the studio site, which adds real friction
If AI2U isn't the right fit
- Want chat depth, not 3D? Candy AI or Anima - the dialogue is better, the character lives in a chat window.
- Want a free indie game in roughly the same shape? Yandere AI Girlfriend Simulator - pay-what-you-want, smaller scope, much darker tone.
- Want to build your own with the same scaffolding? Our DIY guide covers three setups from a 10-minute custom GPT to a self-hosted local LLM with VRM avatar.
- Want voice calls without 3D? EVA AI has the best phone-call experience on the shortlist.
Who should buy AI2U
Buy it if: you already enjoy visual novels or dating sims, you own a PC with at least a GTX 1060 / 4 GB VRAM, you specifically want a one-time purchase instead of a $15/mo subscription, and the prospect of a character whose face actually moves when she's annoyed at you feels more interesting than another chat window.
Skip it if: conversation depth is the only thing you care about (use Candy), you're on a laptop with integrated graphics, you want NSFW out of the box on Steam, or you want to message her from your phone on the train. None of those work here.
AI2U FAQ
The questions we got most after publishing this review.
Can I run AI2U on a laptop?
Only with a dedicated GPU and at least 4 GB of VRAM. Integrated Intel UHD and AMD Vega iGPUs fail the character-load step on launch - the avatar simply doesn't appear. Gaming laptops from 2020 onward (GTX 1060 Mobile and up) generally clear the bar. The Apple Silicon build runs natively on M1/M2/M3/M4 Air and Pro and held 60 fps medium on our M2 Air test.
Is AI2U on consoles?
No, and no console release is planned. PC only - Windows and macOS via Steam. Cloud-streaming services (GeForce Now, Boosteroid) don't carry it as of June 2026, and the studio hasn't signaled it will change.
Is there an NSFW patch?
Yes. Distributed free off-Steam by the studio, gated behind age verification on their site. The patch unlocks adult scenes for specific characters; some DLC packs ship their own optional adult content. The Steam build itself is rated T or M depending on region - the explicit content has to live off-platform because of Steam's category rules for this product type.
Does AI2U really use AI for dialogue, or is it scripted?
Both, deliberately. Cutscene dialogue is scripted and voice-acted. Free-chat hubs and reactive lines within scenes are generated live by a GPT-3.5-class cloud LLM, then routed through per-character TTS tuned to match the cast's tone. The hybrid is why scripted scenes sound clean and free-chat sounds slightly less so.
Will my save file survive updates?
Historically, no. The 2024 1.0 launch and the late-2025 1.5 engine swap both broke older saves. The studio shipped a migration tool in early 2026 and it mostly works, but the failure mode is catastrophic (route progress wiped). Back up %APPDATA%\AI2U\Saves on Windows or ~/Library/Application Support/AI2U on macOS before any patch larger than a hotfix.
How does AI2U compare to a chat app like Candy AI?
Different products in different categories. Candy wins on chat depth, character-locked image generation, voice calls, and dollars-per-month value. AI2U wins on presence - an animated character physically reacting to you in a real-time 3D scene. If you want an ongoing companion app to message daily, use Candy. If you want an evening of being inside a story with a character, AI2U is the only product on our list that delivers it.